CELL article III on RWT

[Gubbi]

Another article on CELL is online at Real World Tech

Cheers
Gubbi

 

[Carl B]

And here I was waiting for Siggraph ray-tracing slides, when all along there was more Realworltech I could have been reading!

Thanks for the heads up Gubbi.

 

[Titanio]

Thanks Gubbi!

Interesting to see him puzzle over the PPE flop rating as we all have at one point or another

Also strange is his commentary on the PPE, the reasons behind its improvement in DD2 etc. He focusses too much, IMO, on the programming model presented by Toshiba at Coolchips (which uses the PPE only for task scheduling) when formulating his discussion on that. Other presentations from SCEA have explained models where the PPE is freed from such tasks to work on computational tasks itself - such usage justifies a boost in PPE vector performance etc.

 

[Carl B]

The gambit to get Apple onboard doesn't seem overly outlandish in and of itself, but having failed to do so, one would think that they would 'drop back down' to the initial version of the Cell, for it's reduced die size and transistor count if nothing else... which is why I'm kind of with Titanio on this one.

So I certainly think there must be some other thinking beyond just a play for Apple at work here; still, it's possible they taped it out and just decided to keep it regardless, afterall, it's not that much more die area to add some decent functionality to the PPE.

Which as we're seeing, is crucial to the flood of next-gen ports.

 

[chachi]

Also strange is his commentary on the PPE, the reasons behind its improvement in DD2 etc. He focusses too much, IMO, on the programming model presented by Toshiba at Coolchips (which uses the PPE only for task scheduling) when formulating his discussion on that. Other presentations from SCEA have explained models where the PPE is freed from such tasks to work on computational tasks itself - such usage justifies a boost in PPE vector performance etc.

The model presented at Coolchips is one you'd use for general computing, this is IMO the original idea behind the processor, where you do all of the work on the individual Cells (SPEs).

On a console where you have explicit control over the hardware resources you can achieve greater efficiency by assigning tasks by hand to the different SPEs and using the PPE as a primary GP core.

I think he's right that Altivec was added to lure Apple but the original version of Cell has it as well. While it would be quite valuable to Apple who has lots of code using Altivec it seems redundant on Cell where you'd want to assign those sorts of tasks to SPE threads. It's hard to say what drove the move to the DD2 design without specifications for either version of the PPE, there was a paper published (404'd since I saw it, but you can see references to it if you google "Waternoose Charlie Johnson") where they talked about using some software they'd to find out design flaws in upcoming hardware and improve efficiency before actually fabbing chips. Waternoose and BBE / Cell were two of the processors that they talked about. It could just be that DD2 is the result of the testing, or that they decided that the original programming model wasn't going to work quite as well as they'd hoped (if you think programmers are complaining *now*, imagine if all their code had to run on SPEs).

 

[Megadrive1988]

Battlelines of the Digital Living Room (A Musing)

In a previous article, this author pointed out that the CELL processor appears to be the cornerstone of STI’s strategy in winning the battle of the digital living room. That is, the CELL processor will form the foundation of Sony’s renewed gambit to bypass the traditional personal computer and capture the digital hub of household communications and entertainment, a lucrative market that is also targeted by Microsoft, Intel, and Apple. Interestingly enough, Microsoft is well positioned regardless of whether the traditional personal computer or the game console wins the battle for the heart of the digital hub. Microsoft can adjust and retain relevance regardless of the outcome of the battle of the digital living room. However, the strategic alliances formed by Microsoft with IBM, and Sony with Toshiba and IBM leaves Intel and Apple out in the cold if the game consoles gain ascendancy as the hub of the digital living room. With the battlelines as given, it seems natural that Apple and Intel would seek to create a third alliance, one that can deliver both the silicon and the software stack to compete for the digital living room against the MI (Microsoft-IBM) alliance and the STI alliance. With the shocking announcements of Apple’s transition from PPC to x86, pundits have espoused various theories and reasons behind the switch. While the short term justifications of the PPC970FX failing to reach 3 GHz, chronic PPC processor shortages from both IBM and Motorola (Freescale), or higher than desired processor prices may all have some validity, the long term justification may be as simple as the formation of the strategic alliance in the battle for the digital living room for players left out by the previous alliances.

A long time ago, before Steve Jobs returned to his role as Apple’s CEO, he was asked what he would do it he was placed in charge of reviving Apple’s flagging fortunes. Steve Job’s answer at the time was that he would keep the Macintosh alive as long as necessary to move on to the next big thing. Following this theme in the years after Steve Job’s return, Apple has positioned itself as the manufacturer of a digital hub where a Macintosh computer is the center of activity for various digital appliances such as a DV camcorder or a music player. In this sense, Sony’s gambit to capture the hub of the digital living room most directly threatens Apple’s future. More ominously, Apple would find that if it were ever pitted against Sony in a battle for the center of the digital hub, the prospect of using the CELL processor in the battle against Sony would be one that is highly unpalatable. That is, if Apple should adopt the CELL processor, it would become a de facto junior partner in the STI alliance. Then, if the battle of the digital living room materializes along the computer-game console battleline, Apple could find itself having to compete against Sony on a processor that Sony co-specified with IBM, a processor that has an unusual programming model with the tool chains being co-developed by Sony, and a processor that Sony plans to manufacture on a massive scale in its own fabrication plant.

The implication is that Apple would find the CELL processor to be undesirable from both technical and strategic perspectives. Steve Job’s ambiguous statements that “Intel has a good roadmap†may be a simple statement that expresses his belief that Intel has a roadmap that is better suited to Apple’s long term strategic plans. However, despite the drawbacks of a CELL-like processor on traditional computer centric applications, it seems clear that a heavily multi-threaded CELL-like approach to multimedia applications is the correct approach. The question then is if Intel will soon embark on or has already secretly embarked on the development of a CELL-like processor that will enable Apple and Intel to challenge the MI and STI alliances. From Apple’s perspective, a roadmap filled with low power mobile processors, cheap Celeron processors, and enhanced with the promise of new types of devices that can enable it to adjust more rapidly to changing market trends would be a roadmap that is ideal to its continued development as a corporate entity and continued battle to win the digital hub of future homes.

oh boy. here we go. Apple and Intel will form the backbone of a new processor alliance, perhaps with additional semiconductor partners (3D graphics and--or CPU partners) to make an effort at a third entertainment platform to take over the living room, to take on PS3/PS4, Xbox2/Xbox3.
say hello to a next-gen Pippin console in the coming years 8) :lol

 

[V3]

Too much speculation in the article to add anything new.

 

[Titanio]

The model presented at Coolchips is one you'd use for general computing, this is IMO the original idea behind the processor, where you do all of the work on the individual Cells (SPEs).

On a console where you have explicit control over the hardware resources you can achieve greater efficiency by assigning tasks by hand to the different SPEs and using the PPE as a primary GP core.

I think the idea outlined in that model was, though, that you wouldn't use the PPE at all for computational tasks, just scheduling for the SPEs. As shown by other models, this needn't be the case, and the PPE can be executing other programmer code aswell (or entirely, with the SPEs looking after themselves).

I think he's right that Altivec was added to lure Apple but the original version of Cell has it as well. While it would be quite valuable to Apple who has lots of code using Altivec it seems redundant on Cell where you'd want to assign those sorts of tasks to SPE threads.

The point is, Altivec/VMX was in the original Cell design too. They simply "beefed it up" (more registers?). Why would it have been there in the first place?

I think the PPE was always meant to be executing general programmer code, or to allow for that..the single focus on the Coolbits model is leading him astray IMO as regards the reasoning for the DD2 revision.

 

[Andy]

oh boy. here we go. Apple and Intel will form the backbone of a new processor alliance, perhaps with additional semiconductor partners (3D graphics and--or CPU partners) to make an effort at a third entertainment platform to take over the living room, to take on PS3/PS4, Xbox2/Xbox3.
say hello to a next-gen Pippin console in the coming years 8) :lol

I think that was a joke more than anything. If you look at the words in brackets next to the title it says "A Musing" or AMUSING, haha very funny : Þ

 

[Rockster]

How does the PPE in DD2 compare to a XeCPU core?

 

[Titanio]

How does the PPE in DD2 compare to a XeCPU core?

MS/IBM hasn't revealed the XeCPU cores, or how exactly they work, but the announced paper spec for both is the same (the flop rating is the exact same). It would appear from articles/dev comments etc. that they are very similar too. There may be some small differences (like perhaps the dot product instruction in XeCores).

 

[chachi]

I think the idea outlined in that model was, though, that you wouldn't use the PPE at all for computational tasks, just scheduling for the SPEs. As shown by other models, this needn't be the case, and the PPE can be executing other programmer code aswell (or entirely, with the SPEs looking after themselves).

The scheduler and the rest of the OS, yeah. The other models seemed to be designed around an interim solution where you're fitting Cell into an already existing workflow.

The point is, Altivec/VMX was in the original Cell design too. They simply "beefed it up" (more registers?). Why would it have been there in the first place?

To placate Apple, or because PPE was a core they'd designed so it could be used for some other purpose (e.g. embedded). Only IBM knows for sure and they ain't talking.

I think the PPE was always meant to be executing general programmer code, or to allow for that..the single focus on the Coolbits model is leading him astray IMO as regards the reasoning for the DD2 revision.

The SPEs represent a large majority of the theoretical processing power in Cell and are in fact the actual embodiment of the "cell" in the design. I don't think they would be meant to be used as a sort of glorified SIMD unit, if you want big things from Cell then you should be moving as much of your work as possible to the SPEs.

On a console you have more control over the low level hardware and so can segment your code and assign it to the different resources appropriately, this is good because it's more efficient but it's not anything you could expect to use in a graphics workstation (the other proposed near-term use for Cell).

So far we've heard developers talk about using the PPE to get their code running and then are looking at the SPEs to supplement that, a better strategy IMO (2nd or 3rd gen maybe?) would be to design your engine around the SPEs and then do what absolutely can't be done on them reasonably well on the PPE. That code might run "better" on the PPE, sure, but there are 7 SPEs available and only one PPE. The problem of course is that's a lot of work that can't be leveraged onto anything else. If they don't do it I don't think you'll see the PS3 games looking any better than the X360 ones, unless RSX is much better than we've been led to believe.

 

[standing ovation]

I've never had an Apple computer before, but from what I've heard, Microsoft has made a ton of money taking Mac OS innovations and making them mainstream with Windows.

But now that both companies are writing software for essentially the same chips, maybe all Apple really wanted was a level playing field?

 

[a688]

I've never had an Apple computer before, but from what I've heard, Microsoft has made a ton of money taking Mac OS innovations and making them mainstream with Windows.

But now that both companies are writing software for essentially the same chips, maybe all Apple really wanted was a level playing field?

Off Topic: I think you heard wrong. Seriously, MS does take designs from others but its nothing new in the industry and its not like Apple isn't guilty of the same thing.

On Topic: Who is it that keeps pushing the Apple + cell rumors?

 

[standing ovation]

So Apple is just as dirty as the rest of them? Wow. That's not something you think about when you think about iPods.

The article that Gubbi provided is pushing the rumors, well, the rationale for them anyway. But there has to be a simpler explanation ...